Section: New Results

Combining Locking and Data Management Interfaces

Participants : Jens Gustedt, Mariem Saied, Daniel Salas.

Handling data consistency in parallel and distributed settings is a challenging task, in particular if we want to allow for an easy to handle asynchronism between tasks. Our publication [1] shows how to produce deadlock-free iterative programs that implement strong overlapping between communication, IO and computation.

An implementation (ORWL) of our ideas of combining control and data management in C has been undertaken, see 6.8. In previous work it has demonstrated its efficiency for a large variety of platforms. In 2016, work on the ORWL model and library has gained vigor with the thesis of Mariem Saied (Inria & University of Strasbourg) and Daniel Salas (INSERM). We also now collaborate on that subject with the TADAAM project team from Inria Bordeaux, where a postdoc has been hired through Inria funding.

In 2016, a new domain specific language (DSL) has been completed that largely eases the implementation of applications with ORWL. In its first version it provides an interface for stencil codes, but extensions towards other types of applications are on their way. The approach allows to describe stencil codes quickly and efficient and leads to substantial speedups, see [14].

In addition, the framework has successfully been applied to encapsulate imaging applications that use certain pipeline patterns to describe dependencies between computational tasks, see [16]. Generally we have been able to use the knowledge of the communication structure of ORWL programs to map tasks to cores and thereby achieve interesting performance gains on multicore architectures, see [20].

In another work we have successfully applied ORWL to process Large Histopathology Images, see [15].